Asemirski / myapi

myapi repo

myapi

This is an example of basic api service using FastAPI and OpenAI. It simply returns answers in Polish language to the provided questions using gpt-4 model knowledge and 'short-term' memory. Short-term memory was implemented to cover course's task needs that is to store conversation data untill question is asked.

Prerequisites

Tools

1. python3
2. Azure Subscription
3. az cli
4. terraform

Local development

Infrastructure

1. Azure Key Vault with a secret called 'openai-api-key'

To run app in Azure cloud

Infrastructure

1. Azure Key Vault with a secret called 'openai-api-key'
2. Linux Web App with python3 runtime and configured access to the Key Vault
3. (Optional) Vnet and service/private endpoints for security.

Infrastructure deployment

To deploy required infrastructure, go to iac folder and execute the following commands:

az login
cd iac
terraform init
terraform apply -var="{IP_ADDRESS_OF_YOUR_MACHINE}" -var="openai_api_key={OPENAI_API_KEY}" -auto-approve

Terraform will deploy all the required infra and you'll need only to deploy the app.

To destroy the infra run the following command:

terraform destroy -var="{IP_ADDRESS_OF_YOUR_MACHINE}" -var="openai_api_key={OPENAI_API_KEY}" -auto-approve

NOTE: Current configuration exposes Key Vault only to web app subnet and to the provided local machine IP address.

Usage

To run service locally execute the following commands:

az login # Required to get access to the key vault
pip3 install -r requirements.txt
python3 main.py 'KEY_VAULT_URL'

Add the command above to the Web App startup command box to run service in a Web App.

Questions should be sent to /api/question endpoint, with the following format:

{
    "question": "What's the capital of Poland"
}

Service returns following response:

{
    "reply": "Stolicą Polski jest Warszawa."
}

In case a piece of an iformation is given:

{
    "question": "My name is Alex"
}

Service saves the info utill a question is asked and returns the following response:

{
    "reply": "Ok. Got it."
}